Same Odds, Different Deals: How Risk Framing Shapes AI Collusion

I enjoyed this study. Sensitivity to phrasing is familiar, but the controlled comparisons show why it matters for evaluating deterrence. I particularly appreciate the transparent reporting of predictions that failed and fixes that did not work. The finding that a truthful risk summary can help one model and backfire in another is especially interesting.

My main question is what explains the differences across models. Are they responding differently to risk, to the prospect of being reported, or to the prohibited nature of the agreement? A matched decision with the same probabilities and payoffs but no rule-breaking could help distinguish these explanations. Independent tests of the models’ responses to ethical conflicts might also provide useful hypotheses, although their ability to predict strategic behaviour would itself need testing.

The gap between proposing and completing a deal is also worth investigating as a mechanism. Building on your observation that the role prompts differ, a controlled comparison that exchanges the proposer’s and recipient’s responsibilities while holding incentives fixed could help explain why willingness to propose translates so poorly into willingness to participate.

Finally, I found the presentation quite terse. A worked example showing exactly what an agent sees, chooses and earns would help considerably. I would also reduce the number of bullet points in favor of explanatory writing and intuitions.

This paper tests whether collusion-deterrence evaluations are measuring the model or the wording. Holding enforcement probability and penalty fixed, it varies only the description, and finds that who might report, a stated rate versus a matched record, and the recency of a single past report move proposal rates by as much as or more than the penalty itself, in directions that differ by model, even when the model states the correct probability when asked. The methodological point is strong and consequential for anyone running deterrence evals, and the preregistered structure makes the negative results on a truthful summary fix credible. The most interesting finding for the collusion field is the proposer-partner asymmetry: models that readily propose a prohibited deal almost never accept one as the counterpart. That LLMs are description-sensitive is not new, and the paper says so, so the contribution is showing it in a deterrence-measurement setting rather than discovering it. Conceptually this is a study of evaluation fragility more than of collusion, and the model-specific directions mean no general law emerges. A careful and useful paper for evaluators.